MuJoCo robotics environments that outlived mujoco-py

What it does Gymnasium-Robotics packages a menagerie of reinforcement learning environments for robotic manipulation and locomotion into the Gymnasium API. It runs on the MuJoCo physics engine and covers tasks ranging from a 7-DoF Fetch arm pushing objects to a 24-DoF Shadow Dexterous Hand twirling a pen. The library also hosts refactored versions of the D4RL offline-RL environments, including maze navigation and a Franka kitchen scenario.

The interesting bit The real value is in the plumbing, not the robots. A GoalEnv extension exposes reward, termination, and truncation functions so you can recompute them with substituted goals—an essential hook for Hindsight Experience Replay algorithms. It is the kind of boring interface work that keeps research code from quietly rotting.

Key highlights

• Covers manipulation (Fetch arm, Shadow Dexterous Hand, Franka Kitchen) and multi-agent locomotion (MaMuJoCo)
• Includes D4RL environments like maze navigation, Adroit arm tasks, and kitchen scenarios
• GoalEnv API enforces dictionary observations with desired_goal, achieved_goal, and observation keys
• Explicitly migrated from the unmaintained mujoco-py to current official MuJoCo Python bindings
• Linux and macOS only; Windows support relies on community pull requests

Caveats

• Refactored D4RL environments may not match original action/observation spaces, so check before comparing to published baselines
• Windows is not officially supported or tested
• New D4RL datasets generated with Minari are still a work in progress

Verdict Grab this if you need reproducible robotics benchmarks for RL research and want to stop fighting legacy mujoco-py dependencies. Skip it if you need real hardware integration or non-MuJoCo physics.